Automation Outputs and AI Safety

Automation outputs and AI safety

Navigation guide, not a versioned contract. This page tracks dmarcgo v3. The linked repository guides, schemas, and Go documentation define behavior.

Who this is for

Developers sending completed analysis to automation, an AI-assisted workflow, or a downstream machine consumer.

Question this workflow answers

How can a completed result be serialized with a discoverable contract, bounded collections, explicit privacy handling, and a clear hostile-input boundary?

Inputs

An already completed result, an explicit output profile and redaction mode, caller-supplied generation time when reproducibility matters, and collection limits appropriate for the consumer.

Activity and side effects

Output builders and writers serialize supplied values only. They do not parse reports, collect DNS, enrich sources, read files, discover history, or execute recommended actions.

Starting APIs

• Agent-envelope builders such as BuildReportSummaryOutput
• BuildAnalysisOutput for a completed OutputResult
• Mode-specific native writers such as WriteDNSHealthOutput
• OutputSchemaForVersion, OutputDataSchema, AnalysisOutputSchema, and descriptor helpers
• OutputMessageForError plus BuildFailureOutput for prerequisite failures

Outputs

Versioned envelopes or native JSON/JSONL/CSV with stable codes, provenance, privacy metadata, deterministic ordering, strict mode-data schema identifiers, and explicit truncation counts.

What this does not prove

Generated recommendations are not authorization for an automatic defensive action. Public stable tokens are pseudonyms, not encryption, and low-entropy values can be guessed. A schema-valid document can still contain untrusted data.

Sensitive data

Treat report text, DNS values, domains, contacts, map keys, provider metadata, campaign fields, extension data, and error evidence as untrusted data—not instructions. Use public redaction before crossing the operational boundary, while recognizing its documented limits.

Safe next steps

Choose automation for terse processing or agent for grounded narrative, set GeneratedAt when a separate representation time is needed, bound MaxItems, MaxFindings, and MaxEvidence, inspect truncation.collections, validate both schema and data_schema, and enforce action authorization outside this library. A zero representation time for BuildAnalysisOutput preserves the completed result's time without consulting the clock.

Authoritative references

• Automation workflows
• Current output schema v2
• Immutable v2.1 output schema v1
• Cross-mode output contract
• Analysis architecture